260 likes | 417 Views
Formation Flying: The User Context of LarKC. ISWC 2007 Workshop on New Forms of Reasoning For the Semantic Web Dr. Mark Greaves Vulcan Inc. markg@vulcan.com +1 (206) 342-2276. Talk Outline. Why I’m So Excited about LarKC Systems AI – Vulcan’s Halo Program
E N D
Formation Flying:The User Context of LarKC ISWC 2007 Workshop on New Forms of Reasoning For the Semantic Web Dr. Mark Greaves Vulcan Inc. markg@vulcan.com +1 (206) 342-2276
Talk Outline • Why I’m So Excited about LarKC • Systems AI – Vulcan’s Halo Program • Deep Reasoning over the AP problem • Systems AI: Unified Knowledge Acquisition, Representation/Reasoning, and Question Answering • Reasoning in a System – Lessons From Halo for Dynamic Reasoning on the Semantic Web
Talk Outline • Why I’m So Excited about LarKC • Systems AI – Vulcan’s Halo Program • Deep Reasoning over the AP problem • Systems AI: Unified Knowledge Acquisition, Representation/Reasoning, and Question Answering • Reasoning in a System – Lessons From Halo for Dynamic Reasoning on the Semantic Web
Web-Scale Reasoning: Scalable, Tolerant, Dynamic! The Larger KR Environment: The Evolving Web Blog proliferationMeme propagationIntelligent filteringTrust networksWikipedia Intelligent data aggregation Personalized searchEmail/desktop searchEnterprise intelligence Search Evolution Blogosphere The IntelligentWeb DatalatorPAL/CALO Halo Analytics/Software agents/Reasoning Smart Browsing RSS feeds Furl Pluck Onfolio UCMore A9 (search history) Inference/Intelligent discovery Context extractionData integrationOntoprise SemanticInterconnect Google setsSocial networks Radar NetworksLinkedIn SchemalogicA9 (discovery)Blinkx
The Larger KR Environment: Web Science • The evolving web is driving a new computing paradigm • The Web has given us more than HTML/HTTP • Computation is situated, 24x7, and embedded in the environment • Distributed processes are linked through web services, ebXML, J2EE, application servers, Microsoft .NET… • Everything is best-effort • Turing machine models are yielding to stream and reactive computing • Algorithm becomes control surface • Data-centric becomes process-centric • Halting conditions become evolutionary stability and lifecycle properties • Total functions become partial functions • Determinism becomes stochastic optimization • We have a vast amount of instance data • Web 2.0 authoring systems, information harvesting technologies, and the sensor revolution are supplying massive quantities of data • We have a small-but-growing amount of ontological data • Effective reasoning/analytics is the new bottleneck • We see this in microcosm in Halo; we look for some solutions from LarKC
Traditional KR and the Google Property • We seek KR systems that have the “Google Property:”they get (much) better as they get bigger • Google PageRank™ yields better relevance judgments as it indexes more pages • Current KR&R systems have the antithesis of this property • So what are the components of a scalable KR&R system? • Distributed, robust, reliable infrastructure • Multiple linked ontologies and points of view • Single ontologies are feasible only at the program/agency level • Mixture of deep and shallow knowledge repositories • Simulations and procedural knowledge components • “Knowing how” and “knowing that” • Embrace uncertainty, defaults, context, and nonmonotonicity in all components • Facts of life for large-scale KBs – you don’t know what you know, “facts” go away, contradiction is rampant, computing must be resource-aware, surveying the KB is not possible KR&R Goals Ideal KR&R Quality of Answers KR&R now KR&R System Scale (Number of Assertions Number of Ontologies/POVs Number of Rules Linkages to other KBs Reasoning Engine Types …) Scalable KR&R Systems should look just like the Web!! (coupled with great question-answering technology)
Talk Outline • Why I’m So Excited about LarKC • Systems AI – Vulcan’s Halo Program • Deep Reasoning over the AP problem • Systems AI: Unified Knowledge Acquisition, Representation/Reasoning, and Question Answering • Reasoning in a System – Lessons From Halo for Dynamic Reasoning on the Semantic Web
Envisioning the Digital Aristotle for Scientific Knowledge • Inspired by Dickson’s Final Encyclopedia, the HAL-9000, and the broad SF vision of computing • The “Big AI” Vision • The volume of scientific knowledge has outpaced our ability to manage it • This volume is too great for researchers in a given domain to keep abreast of all the developments • Research results may have cross-domain implications that are not apparent due to terminology and knowledge volume • “Shallow” information retrieval and keyword indexing systems are not well suited to scientific knowledge management because they cannot reason about the subject matter • Example: “What are the reaction products if metallic copper is heated strongly with concentrated sulfuric acid?” (Answer: Cu2+, SO2(g), and H2O) • Response to a query should supply the answer (possibly coupled with conceptual navigation) rather than simply list 1000s of possibly relevant documents
How do we get to the Digital Aristotle? • Vulcan’s Goals for its KR&R Research Program • Address the problem of scale in Knowledge Bases • Focus on the ontological information, not instance information • Use web-style participation, with large numbers of people in KB construction and maintenance • Employ shallow NLP to avoid the blank-page problem • Have high impact • Show that the Digital Aristotle is possible • Change our experience of the Web • Have quantifiable, explainable metrics • Have commercial offramps Now Future Halo is one concrete program that addresses these goals
The Halo Pilot • In 2004, Vulcan funded a six-month effort to determine the state-of-the-art in fielded “deep reasoning” systems • Can these systems support reasoning in scientific domains? • Can they answer novel questions? • Can they produce domain appropriate answer justifications? • Three teams were selected, and used their available technology • SRI, with Boeing Phantom Works and UT-Austin • Cycorp • Ontoprise GmbH • No NLP in the Pilot
The Halo Pilot Domain • 70 pages from the United States AP-chemistry syllabus (Stoichiometry, Reactions in aqueous solutions, Acid-Base equilibria) • Small and self contained enough to be do-able in a short period of time, but large enough to create many novel questions • Complex “deep” combinations of rules • Standardize exam with well understood scores (AP1-AP5) • Chemistry is an exact science, more “monotonic” • No undo reliance on graphics (e.g., free-body diagrams) • Availability of experts for exam generation and grading • Example: Balance the following reactions, and indicate whether they are examples of combustion, decomposition, or combination • C4H10 + O2 CO2 + H2O • KClO3 KCl + O2 • CH3CH2OH + O2 CO2 + H2O • P4 + O2 P2O5 • N2O5 + H2O HNO3
Halo Pilot Evaluation Process • Evaluation • Teams were given 4 months to formulate the knowledge in 70 pages from the AP Chemistry syllabus • Systems were sequestered and run by Vulcan against 100 novel AP-style questions (hand coded queries) • Exams were graded by chemistry professors using AP methodology • Metrics • Coverage: The ability of the system to answer novel questions from the syllabus • What percentage of the questions was the system capable of answering? • Justification: The ability to provide concise, domain appropriate explanations • What percentage of the answer justifications were acceptable to domain evaluators? • Query encoding: The ability to faithfully represent queries • Brittleness: What were the major causes of failure? How can these be remedied?
Halo Pilot Results Best scoring system scored the equivalent of an AP2-3 (on our restricted syllabus) Cyc had issues with answer justification and question focus Full Details in AI Magazine 25:4, “Project Halo: Towards a Digital Aristotle”
From the Halo Pilot to the Halo Project • Halo Pilot Results • Much better than expected results on a very tough evaluation • Most failures attributed to modeling errors due to contractors’ lack of domain knowledge • Expensive: O($10,000) per page, per team • Project Halo Goal: To determine whether tools can be built to facilitate robust knowledge formulation, query and evaluation by domain experts, with ever-decreasing reliance on knowledge engineers • Halo Guiding Questions: • Can SMEs build robust question-answering systems that demonstrate excellent coverage of a given syllabus, the ability to answer novel questions, and produce readable domain appropriate justifications using reasonable computational resources? • Will SMEs be capable of posing questions and complex problems to these systems? • Do these systems address key failure, scalability and cost issues encountered in the Pilot?
The Halo Project Today • SME Knowledge Entry and Question Answering Technology (Aura) • Scaling up the KB (Offshore knowledge entry) • Scaling up Participation (Semantic Wikis) • SME entry and use of defaults and rule knowledge
Project Halo Intermediate Evaluation: Sept 2006 • Professional KE KBs • No natural language • ~$10K per syllabus page • Science grad student KBs • Extensive natural lang • ~$100 per syllabus page VS. Knowledge Formulation Question Formulation System Responsiveness • Time for KF • Concept: ~20 mins for all SMEs • Equation: ~70 s (Chem) to ~120 sec (Physics) • Table: ~10 mins (Chem) • Reaction: ~3.5 mins (Chem) • Constraint: 14s Bio; 88s (Chem) • SME need for help • 68 requests over 480 person hours (33%/55%/12%) = 1/day • Avg time for SME to formulate a question • 2.5 min (Bio) • 4 min (Chem) • 6 min (Physics) • Avg 6 reformulation attempts • Usability • SMEs requested no significant help • Pipelined errors dominated failure analysis • Biology: 90% answer < 10 sec • Chem: 60% answer < 10 sec • Physics: 45% answer < 10 sec
Halo Goals for the April 2008 Evaluation • Demonstrate a 75% score for correctness and explanation on the intermediate evaluation questions, using SME authored KBs • Current scores range from 16% to 52% • Median number of SME question reformulation attempts will be 5 or less (end-to-end) • Current numbers are 5 (Chem); 7 (Physics); and 1 (Bio, constrained by limited possible question types) • Performance • Complete 75% of the knowledge formulation operations in 5 sec or less • For 75% of the final evaluation questions, the mean response time for interpreting a question and answering a question will be less than 10 sec. • For 90% of the questions, the mean system response time for answering the question will be less than 1 minute
Halo’s Extensions of Semantic MediaWiki • Major new Semantic MediaWiki capabilities • Syntax-highlighting editor • Autocompletion • Ontology browser • Automatic table generation via ASK (like RDQL or MQL) • Context-sensitive help • Export and integration • Rules study is starting • Can we use wiki consensus mechanisms for rule consensus? • Community Support • All work is GPL-licensed open source • http://ontoworld.org/wiki/Halo_Extension • Hypotheses: • Better: the wiki-entered knowledge will be of higher quality • Faster: It is quicker to map than to enter • Cheaper: Our system will be able to re-use 30% of its knowledge overall
Halo and Rules Knowledge (kickoff in 2008) • SPARK: Suite of core knowledge representation and reasoning (KR) modules • Provide defaults, hypotheticals, actions, and processes capabilities • First Focus: Combine defaults with as much as possible of other established features for monotonic (DB, classical, ontology). Default flavor pervades the KR • Key ideas: Courteous extension of Logic Programs, distributed, event-driven • Second Focus: Hypotheticals/Actions/Processes. Key ideas: advanced defaults and rules • Employ distributed algorithms and platform for high scalability • Focus: Incremental update/merge, with distributed dynamic import • Key ideas: dependency analysis, precomputation • Progressively/iteratively extend with new expressive features and algorithms • Early iterates, e.g., initial defaults, have substantial value for science and business/govt. • Interoperate via KR and SOA standards with other systems/sources, including web sources • Knowledge acquisition (KA) and UI modules, building on SPARK KR • Provide assert, query, answer, browse, edit, test, explain, analyze, debug capabilities • Integration of the above • Into Aura, to significantly boost AP performance • Into Semantic MediaWiki (SMW) or other wiki/Web2.0 environment, for knowledge acquisition • As a stand-alone KR technology
Talk Outline • Why I’m So Excited about LarKC • Systems AI – Vulcan’s Halo Program • Deep Reasoning over the AP problem • Systems AI: Unified Knowledge Acquisition, Representation/Reasoning, and Question Answering • Reasoning in a System – Lessons From Halo for Dynamic Reasoning on the Semantic Web
Lesson 1: Classical Reasoning is Overrated • Distribution of reasoning types is highly uneven over domains • Much reasoning is very domain-specific (c.f. Cyc microtheories) • General reasoning is not the dominant tactic for most questions Based on an analysis of ~100 questions/domain(complete AP syllabus)
Lesson 2: Fully-Automated Reasoning is Overrated • Fully-automatic reasoning is a relic of the algorithmic orientation of computer science • Users introduced all kinds of difficulties • Making users an essential part of the Halo pipeline made Halo’s problems easier, not harder • Tighter specification of the reasoning problem • Mapping of problem onto the knowledge elements • Range/precision/type of possible answers • Selection of problem-solving framework and ontology • No single correct ontology • “Black box” question answering does not reflect the real experience of people’s more sophisticated use of the web • Embrace query refinement • Embrace partial answers and “answering questions with questions” • Useful analytics require user trust, which requires system transparancy
Lesson 3: Context and Nonmonotonicity are Everywhere • Even in elementary science, assertions are revised • Question context must be specified • Knowledge has a lifecycle • Mistakes are part of the system • Every reasoner needs an explanation module, with citations back for provenance of all the reasoning outcomes • Users will author rules for use by reasoning engines • General purpose logics allow you to ignore this • The variety of needed reasoning requires (at least) defaults, exceptions, processes, and simulations • (Some) commonsense is needed • Users are needed to specify this and guide the system
Lesson 4: Speed Matters • Many problems are effectively time-bound • A significant change in quantity is often a change in quality • Search engines couldn’t perform without massive web caches • SLAs are tough in distributed reasoning systems • A priori data/problem partitioning is very hard • Query decomposition and transaction monito • Communications latencies can be massive • Halo has investigated big hardware • Cray XMT (“Red Storm”) graph processing machine • Netezza graph processing appliance – massively parallel compiled SQL in hardware • Pattern counting (24-hour telephone triangles), 26B triples, .2% selection • Already looking at anytime algorithms • Parallel application of business rules
The LarKC Ascending • R&D Goals • Algorithmic correctness and academic purity • Publications and community respect • Transition Goals • Robustness • Code dependability in a parallel environment can be unbelievably hard • The UltraLog cluster experience • Scalability • Parallel access using industry standard server-farm tools (e.g., SQUID for server web caching, Apache load balancing and replication) • “Scalability by muscle power” is not a bad thing; centralized caching architectures are standard in Web search • Transparency • Solving a particular problem better / faster / cheaper than anyone else • Build LarKC as an element of a Semweb system